Protective-action decisions in the first hours of a nuclear power plant accident, such as sheltering, evacuation and iodine prophylaxis, depend on atmospheric dispersion forecasts, which in turn depend on an estimate of the source term. When in-plant instrumentation is lost, as at Chernobyl and Fukushima Daiichi, the source term must be reconstructed from off-site measurements, and classical inversion methods that need thousands of forward-model runs are often too slow for real-time decision support. This study assesses two structurally different neural networks, a Gated Recurrent Unit (GRU) and a Patch Time Series Transformer (PatchTST), as fast surrogates that forecast the release rate over the next six hours from a five-sensor downwind network. A corrected Gaussian plume model (GPM+) with wind-profile correction, Pasquill–Gifford stability classes, wet and dry deposition and radioactive decay generates 16,000 synthetic release scenarios per case. Each machine learning model receives 14 input features and is tested with a 24 hr and a 48 hr input window, with hyperparameters tuned by Optuna and results averaged over three seeds. With the 48 hr window, the two models are equally accurate (Log10RMSE 0.212 for GRU, 0.215 for PatchTST), but PatchTST has 31% fewer parameters and about 8.4 times lower CPU latency. With the 24 hr window GRU is more accurate (0.153 versus 0.173) but larger and slower. Both models outperform an idealized persistence baseline, which is given the true simulated release rate rather than an estimate of it, from the t+3 horizon onward, and 90.4–94.1% of forecasts fall within a factor of two of the true release rate. Because both networks require a complete input window before they can issue a forecast, the first forecast is only available 24 or 48 hr after monitoring begins; the results indicate how model choice can be matched to the accuracy, computing and timing constraints of an emergency decision-support system.

暂无